What a language model actually stores
A language model is often described as if it contained a library: facts on shelves, ready to be fetched. That picture is convenient and misleading. What the system stores, after training, is a very large set of numerical parameters. Those parameters make one operation likely: given some text, propose a continuation. This page stays with that operation. It does not say whether any product is useful, and it does not turn the description into a market view.
Invest Star Assets files this under Models. The reader is someone who meets phrases such as “the model knows” in articles about technology and, sometimes, in commentary about companies and markets. The phrase needs a narrower meaning before it is repeated.
Prediction, not retrieval
During training, the system sees enormous amounts of text and adjusts its parameters so that a likely next piece of text scores higher than an unlikely one. The result is not an index you can open at a page. Ask for a definition and the model produces a sentence that resembles definitions it was shaped by. Ask for a number and it produces digits that resemble numbers in similar contexts. Resemblance is not a lookup.
Retrieval would imply a stored record and a way to point at it. A base language model does not point. Some products later attach search, a database, or a citation step. Those attachments are extra systems. They are not what the parameters themselves are. When a paragraph says “the AI retrieved the fact,” ask which extra system, if any, was in the path.
- Parameter
- A number inside the model that was adjusted during training. It is not a sentence.
- Continuation
- The text the model assigns a high score to, given what came before.
- Record
- A document you can cite. The model does not keep one unless another system supplies it.
Weights are not sentences
People sometimes imagine that a famous paragraph sits inside the model, intact, waiting. Training does not file paragraphs that way. It compresses regularities: how words tend to follow other words, how a definition tends to be shaped, how a cautious hedge tends to appear. A specific sentence can be approximated if it was common enough to leave a trace. Approximation is still not storage of the sentence as a quotable source.
This is why two runs can differ, and why a model can sound sure about a detail that was never stable in the training text. The parameters encode tendency. Tendency does not carry a footnote. If you need the original wording, you need the original document, not a regeneration of its style.
What a prompt adds
A prompt is the text you place in front of the model at the moment of use. It is not part of the stored parameters, except in products that later save it. The prompt steers which continuation is likely: a formal definition, a short list, a refusal. Steering is not the same as checking. A well-shaped prompt can still be followed by a fluent error, because the scoring operation does not consult the world.
Instructions such as “only use reliable sources” are themselves just more text. Unless a separate tool actually opens those sources and the answer is tied to them, the instruction changes style more than evidence. Style can look like caution. Caution in tone is not a citation.
What is absent
A model does not store your private files unless a product is built to upload them. It does not store tomorrow’s prices, a regulator’s next sentence, or a guarantee about any asset. It also does not store a verdict on whether a technology company is a sound holding. Those are different questions, and this catalog does not answer them.
What is present is a machine for likely language. That is enough to explain why an answer can be useful as a draft of wording and still be unfit as a source. The next filing, on training data, asks where the likelihood came from. This page stops at the storage claim: parameters, not a library.
- Say “produces a continuation,” not “knows.”
- Separate the model from any search tool attached later.
- If you need a quote, open the document.
Close
A language model stores adjusted numbers that score continuations. It does not store a checked archive, and it does not store advice. Treating the output as a retrieved fact skips the only mechanism the system actually has.